SonarPowerNoise (#2)
* SonarPowerNoise: WIP * don't allow highpass > lowpass * fix filter unit gain The filter was computed as an amplitude gain map. This change computes energy gains instead. This makes easier to ensure unit-gain and makes the alpha values more in line with literature where the power spectrum is obeying the power law. new alpha = old alpha / 2 Also fixes the gain of common_mode. * allow stretch < 1.0, paramter range fixes * rename lowpass/highpass to min_freq/max_freq * remove torch.no_grad wrappers * add previews * static seed for previews * README: SonarPowerNoise documentation
This commit is contained in:
elias-gaeros
@@ -70,6 +70,30 @@ The sampler and `NoisyLatentLike` nodes now take an optional `SonarCustomNoise`
|
||||
|
||||
**Note**: If you connect the optional `SonarCustomNoise` node to a Sonar sampler, the `NoisyLatentLike` node or the `SamplerConfigOverride` node, it will override the noise type selected in the node.
|
||||
|
||||
### SonarPowerNoise
|
||||
|
||||
The `SonarPowerNoise` node generates [fractional Brownian motion (fBm) noise](https://en.wikipedia.org/wiki/Fractional_Brownian_motion#Frequency-domain_interpretation). It offers versatility in producing various types of noise including gaussian, pink, 2D brownian noise, and all intermediates.
|
||||
|
||||
By default, the node generates normal gaussian noise. Here's an overview of its parameters:
|
||||
|
||||
- `factor` and `rescale` operate similarly to `SonarCustomNoise`, enabling the addition of multiple sources of noises.
|
||||
- `time_brownian` introduces correlation across sampler timesteps for SDE solvers.
|
||||
- `alpha` is the main parameter. `alpha > 0` amplifies low frequencies; `alpha = 1` yields pink noise, and `alpha = 2` produces brownian noise. Conversely, for `alpha < 0`, it amplifies high frequencies.
|
||||
- `min_freq` and `max_freq` determine the range of frequencies allowed through. Setting `max_freq = `$\sqrt{1/2} \simeq 0.7071$ enables the passage of the highest frequencies. In cases where `alpha < 0`, setting `max_freq = 0.5` is advisable to diminish the power of diagonally oriented frequencies.
|
||||
- `stretch`, `rotate`, and `pnorm` alter the filter's shape by stretching, rotating, or cushioning the band-pass region.
|
||||
- Lowering `mix` moderates the filter's effect by blending back unfiltered gaussian noise from the same sample.
|
||||
- `common_mode` is an attempt to desaturate the latent by injecting the average across channels into every latent channel. However, this may result in a specific color due to the encoding of the unit vector by the latent space. Note that this is done _after_ the `mix`ing of unfiltered gaussian noise.
|
||||
- Enabling `preview` provides a visual representation of the filter. `no_mix` sets `mix = 1` for the preview. The preview includes, from left to right:
|
||||
- Fourier domain visualization: Low frequencies at the center, with black indicating filtered-out frequencies.
|
||||
- Spatial visualization of the 2D kernel: The filtering can be interpreted as convolution with the displayed kernel.
|
||||
- Sample: Gaussian sample with shaped frequency spectrum. A single latent channel will look like this.
|
||||
|
||||
**Frequency-domain Interpretation**: The Fourier transform decomposes a 2D latent into sinusoids covering all spatial orientations and frequencies. For an independent and identically distributed gaussian sample, energy is evenly distributed across all frequencies and orientations. Scaling the power spectrum by $1 / f^\alpha$, where $\alpha>0$, boosts low frequencies, introducing spatial correlations.
|
||||
|
||||
**Spatial Domain Interpretation**: A gaussian latent sample comprises independently sampled pixels, exhibiting no spatial correlations. Conversely, a requirement that each pixel value differs from its neighbors by a $\epsilon \sim \mathcal{N}(0, 1)$ results in 2D brownian noise ($\alpha=2$).
|
||||
|
||||
**Seed Considerations**: While the node defaults to outputting gaussian noise, a given seed produce a different sample than the one produced by other gaussian noise sources. This stems from sampling the noise directly in the frequency domain to avoid the cost of a FFT. When `time_brownian = true`, noise sampling occurs in the spatial domain, ensuring that default parameters yield output equivalent to `SonarCustomNoise` set to `brownian`.
|
||||
|
||||
## Related
|
||||
|
||||
I also have some other ComfyUI nodes here: https://github.com/blepping/ComfyUI-bleh/
|
||||
|
||||
+2
-1
@@ -1,4 +1,4 @@
|
||||
from .py import nodes, sonar
|
||||
from .py import nodes, powernoise, sonar
|
||||
|
||||
sonar.add_samplers()
|
||||
|
||||
@@ -9,6 +9,7 @@ NODE_CLASS_MAPPINGS = {
|
||||
"SamplerConfigOverride": nodes.SamplerNodeConfigOverride,
|
||||
"NoisyLatentLike": nodes.NoisyLatentLikeNode,
|
||||
"SonarCustomNoise": nodes.SonarCustomNoiseNode,
|
||||
"SonarPowerNoise": powernoise.SonarPowerNoiseNode,
|
||||
"SonarGuidanceConfig": nodes.GuidanceConfigNode,
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,333 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import math
|
||||
import os
|
||||
import random
|
||||
|
||||
import folder_paths
|
||||
import torch
|
||||
from comfy.k_diffusion.sampling import BrownianTreeNoiseSampler
|
||||
from PIL import Image
|
||||
from torch import Tensor
|
||||
|
||||
from .nodes import SonarCustomNoiseNodeBase
|
||||
from .noise import CustomNoiseItemBase
|
||||
|
||||
# ruff: noqa: ANN003, FBT001, FBT002
|
||||
|
||||
|
||||
class PowerNoiseItem(CustomNoiseItemBase):
|
||||
def __init__(self, factor, **kwargs):
|
||||
super().__init__(factor, **kwargs)
|
||||
self.max_freq = max(self.max_freq, self.min_freq)
|
||||
|
||||
def make_filter(self, shape, oversample=4, rel_bw=0.125):
|
||||
"""Construct a band-pass * 1/f^alpha filter in rfft space."""
|
||||
height, width = shape[-2:]
|
||||
hfreq_bins = width // 2 + 1
|
||||
|
||||
# Flat unit gain frequency response
|
||||
if self.mix < 1.0:
|
||||
flat = torch.ones(1, 1, height, hfreq_bins)
|
||||
if self.mix <= 0.0:
|
||||
return flat
|
||||
|
||||
# Start with an over-sampled fftshift(rfft2freq()) grid. uses complex
|
||||
# numbers for convenient 2d rotation (unrelated to the fft complex phase
|
||||
# space)
|
||||
fc = torch.complex(
|
||||
# real-fftfreq
|
||||
torch.linspace(0, 0.5, oversample * hfreq_bins),
|
||||
# normal fftfreq
|
||||
torch.linspace(
|
||||
-(height // 2) / height,
|
||||
((height - 1) // 2) / height,
|
||||
oversample * height,
|
||||
).unsqueeze(1),
|
||||
)
|
||||
# Rotate, stretch and p-norm
|
||||
if abs(self.rotate) >= 1e-3:
|
||||
fc *= torch.exp(1.0j * torch.deg2rad(torch.scalar_tensor(self.rotate)))
|
||||
if self.stretch > 1.0:
|
||||
fc.real *= self.stretch
|
||||
else:
|
||||
fc.imag *= 1.0 / self.stretch
|
||||
if abs(self.pnorm - 2.0) < 1e-3:
|
||||
d = fc.abs()
|
||||
else:
|
||||
d = (
|
||||
torch.view_as_real(fc)
|
||||
.abs()
|
||||
.pow(self.pnorm)
|
||||
.sum(-1)
|
||||
.pow(1.0 / self.pnorm)
|
||||
)
|
||||
|
||||
# filter gain function
|
||||
op = torch.empty_like(d)
|
||||
m_highpass = d >= self.min_freq
|
||||
m_lowpass = d < self.max_freq
|
||||
m_band = m_highpass & m_lowpass
|
||||
# 1 / f^alpha for the band-pass region
|
||||
op[m_band] = d[m_band].pow(-self.alpha)
|
||||
# easing gaussian (TODO: try cosine windows)
|
||||
m_lowpass = ~m_lowpass
|
||||
op[m_lowpass] = math.pow(self.max_freq, -self.alpha) * torch.exp(
|
||||
-(d[m_lowpass] - self.max_freq).square() / (rel_bw * self.max_freq) ** 2,
|
||||
)
|
||||
if self.min_freq > 0.0:
|
||||
m_highpass = ~m_highpass
|
||||
op[m_highpass] = math.pow(self.min_freq, -self.alpha) * torch.exp(
|
||||
-(d[m_highpass] - self.min_freq).square()
|
||||
/ (rel_bw * self.min_freq) ** 2,
|
||||
)
|
||||
op = torch.nn.functional.interpolate(
|
||||
op[None, None, ...],
|
||||
(height, hfreq_bins),
|
||||
mode="bilinear",
|
||||
align_corners=True,
|
||||
)
|
||||
op = op.roll(-(height // 2), -2) # ifftshift
|
||||
if self.alpha > 0:
|
||||
# In general, the mean offset should be kept as is, sampled from
|
||||
# N(0, 1 / sqrt(H*W) ). However, gain goes to inf when alpha>0.
|
||||
op[..., 0, 0] = 0
|
||||
|
||||
# Scale to unit power gain, then mix flat filter
|
||||
mean_pow_gain = op.mean()
|
||||
if mean_pow_gain <= 0.0:
|
||||
# don't fail catastrophically when something broke
|
||||
return flat
|
||||
op *= 1.0 / mean_pow_gain
|
||||
if self.mix < 1.0:
|
||||
op = torch.lerp(flat, op, self.mix, out=op)
|
||||
return op.sqrt_()
|
||||
|
||||
def make_noise_sampler(
|
||||
self,
|
||||
x: Tensor,
|
||||
sigma_min: float | None,
|
||||
sigma_max: float | None,
|
||||
seed: int | None,
|
||||
cpu: bool = True,
|
||||
):
|
||||
shape = x.shape
|
||||
device = x.device
|
||||
time_brownian = self.time_brownian
|
||||
if self.time_brownian:
|
||||
if sigma_min is None:
|
||||
raise ValueError(
|
||||
"time correlated brownian mode is valid only for stochastic samplers",
|
||||
)
|
||||
brownian_tree = BrownianTreeNoiseSampler(
|
||||
x,
|
||||
sigma_min,
|
||||
sigma_max,
|
||||
seed=seed,
|
||||
cpu=cpu,
|
||||
)
|
||||
|
||||
common_mode = self.common_mode
|
||||
if common_mode > 0.0:
|
||||
b, c, h, w = shape
|
||||
torch.eye(c, c)
|
||||
channel_mixer = torch.lerp(
|
||||
torch.eye(c, c),
|
||||
torch.ones(c, c) / c,
|
||||
common_mode,
|
||||
)
|
||||
channel_mixer = channel_mixer.sqrt().to(device, non_blocking=True)
|
||||
|
||||
filter_rfft = self.make_filter(shape).to(device, non_blocking=True)
|
||||
|
||||
def sampler(sigma, sigma_next):
|
||||
if time_brownian:
|
||||
noise = brownian_tree(sigma, sigma_next).to(device)
|
||||
noise_rfft = torch.fft.rfft2(noise, norm="ortho")
|
||||
else:
|
||||
noise_rfft = torch.randn(
|
||||
(*shape[:-1], filter_rfft.shape[-1]),
|
||||
dtype=torch.complex64,
|
||||
device=device,
|
||||
)
|
||||
noise = torch.fft.irfft2(
|
||||
noise_rfft.mul_(filter_rfft),
|
||||
s=shape[-2:],
|
||||
norm="ortho",
|
||||
)
|
||||
|
||||
if common_mode > 0.0:
|
||||
noise = channel_mixer @ noise.swapaxes(0, 1).reshape(c, -1)
|
||||
noise = noise.reshape(c, b, h, w).swapaxes(1, 0)
|
||||
return noise.mul_(self.factor)
|
||||
|
||||
return sampler
|
||||
|
||||
def preview(self, size=(128, 128)):
|
||||
filter_rfft = self.make_filter(size, oversample=1)
|
||||
filter_fft = rfft2_to_fft2(filter_rfft)
|
||||
noise = torch.fft.irfft2(
|
||||
filter_rfft
|
||||
* torch.randn(
|
||||
filter_rfft.shape,
|
||||
dtype=torch.complex64,
|
||||
generator=torch.Generator().manual_seed(0),
|
||||
),
|
||||
s=size,
|
||||
norm="ortho",
|
||||
)
|
||||
kernel = torch.fft.irfft2(filter_rfft, s=size, norm="ortho")
|
||||
kernel = kernel.roll((size[0] // 2, size[1] // 2), (-2, -1))
|
||||
img = (
|
||||
torch.cat(
|
||||
[
|
||||
filter_fft.mul_(1 / 3).tanh_().mul_(256.0),
|
||||
kernel.mul_(1 / 3).tanh_().add_(1.0).mul_(128.0),
|
||||
noise.mul_(1 / 3).tanh_().add_(1.0).mul_(128.0),
|
||||
],
|
||||
dim=-1,
|
||||
)
|
||||
.clamp(0, 255)
|
||||
.to(torch.uint8)
|
||||
)
|
||||
return Image.fromarray(img[0, 0].numpy())
|
||||
|
||||
|
||||
def rfft2_to_fft2(x):
|
||||
"""Apply hermitian-summetry to reconstruct the second half of a fft.
|
||||
|
||||
Only for previews.
|
||||
"""
|
||||
height, width = x.shape[-2:]
|
||||
x_r = x.roll(height // 2, -2) # torch.fft.fftshift(x, -2)
|
||||
x_l = x_r[..., 1 : -1 if width & 1 else None]
|
||||
x_l = torch.flip(x_l.conj(), dims=(-2, -1))
|
||||
if height & 1 == 0:
|
||||
x_l = x_l.roll(1, -2)
|
||||
return torch.cat([x_l, x_r], dim=-1)
|
||||
|
||||
|
||||
class SonarPowerNoiseNode(SonarCustomNoiseNodeBase):
|
||||
@classmethod
|
||||
def INPUT_TYPES(cls):
|
||||
result = super().INPUT_TYPES()
|
||||
result["required"] |= {
|
||||
"time_brownian": ("BOOLEAN", {"default": False}),
|
||||
"alpha": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 0.0,
|
||||
"min": -5.0,
|
||||
"max": 5.0,
|
||||
"step": 0.001,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"max_freq": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 0.7071,
|
||||
"min": 0.0,
|
||||
"max": 0.7071,
|
||||
"step": 0.001,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"min_freq": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 0.0,
|
||||
"min": 0.0,
|
||||
"max": 0.7071,
|
||||
"step": 0.001,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"stretch": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 1.0,
|
||||
"min": 0.01,
|
||||
"max": 100,
|
||||
"step": 0.1,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"rotate": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 0,
|
||||
"min": -90,
|
||||
"max": 90,
|
||||
"step": 5,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"pnorm": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 2,
|
||||
"min": 0.125,
|
||||
"max": 100,
|
||||
"step": 0.1,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"mix": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 1.0,
|
||||
"min": 0.0,
|
||||
"max": 1.0,
|
||||
"step": 0.001,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"common_mode": (
|
||||
"FLOAT",
|
||||
{
|
||||
"default": 0.0,
|
||||
"min": 0.0,
|
||||
"max": 1.0,
|
||||
"step": 0.001,
|
||||
"round": False,
|
||||
},
|
||||
),
|
||||
"preview": (["none", "no_mix", "mix"],),
|
||||
}
|
||||
return result
|
||||
|
||||
def get_item_class(self):
|
||||
return PowerNoiseItem
|
||||
|
||||
def go(
|
||||
self,
|
||||
preview="none",
|
||||
**kwargs,
|
||||
):
|
||||
result = super().go(**kwargs)
|
||||
if preview == "none":
|
||||
return result
|
||||
if preview == "no_mix":
|
||||
kwargs["mix"] = 1.0
|
||||
img = PowerNoiseItem(**kwargs).preview()
|
||||
|
||||
output_dir = folder_paths.get_temp_directory()
|
||||
prefix_append = "sonar_temp_" + "".join(
|
||||
random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5) # noqa: S311
|
||||
)
|
||||
full_output_folder, filename, counter, subfolder, _ = (
|
||||
folder_paths.get_save_image_path(prefix_append, output_dir)
|
||||
)
|
||||
filename = f"{filename}_{counter:05}_.png"
|
||||
file_path = os.path.join(full_output_folder, filename) # noqa: PTH118
|
||||
img.save(file_path, compress_level=1)
|
||||
|
||||
return {
|
||||
"ui": {
|
||||
"images": [
|
||||
{"filename": filename, "subfolder": subfolder, "type": "temp"},
|
||||
],
|
||||
},
|
||||
"result": result,
|
||||
}
|
||||
Reference in New Issue
Block a user